Integração da tecnologia CUDA ao modelo de previsão do tempo Eta

• Ano de defesa: 2018
• Autor(a) principal: Flores, Henrique Gavioli
• Orientador(a): Rebonatto, Marcelo Trindade
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade de Passo Fundo
• Programa de Pós-Graduação: Programa de Pós-Graduação em Computação Aplicada
• Departamento: Instituto de Ciências Exatas e Geociências – ICEG
• País: Brasil
• Palavras-chave em Português: Arquitetura de computador; Meteorologia; Programação paralela (Computação); Programas de computador
• Área do conhecimento CNPq: CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
• Link de acesso: http://tede.upf.br/jspui/handle/tede/1466

Objective: The main objective of this dissertation is to implement the CUDA technology for data processing in the Eta weather model in order to reduce its execution time. Methods and Materials: A study on the prediction models and the Fortran language was presented, which was used for the writing of the Eta. The Eta model processing was studied and mapped, indicating that the data processing is done by dividing the MPI tasks into two groups: I / O servers, responsible for storing the generated data, and forecasting processes, which compute each part of the area to be predicted. A code analysis was performed to identify computationally intensive activities, in order to identify points if more than one level of parallelism could be applied. After defining a two-level parallelism model, CUDA technology was chosen to try to obtain a performance gain without changing the logic used by the model. Results: The model was implemented using CUDA, being chosen four points of greater complexity of code and execution time to implement the second level of parallelism. Twenty replicates were performed for various combinations of process numbers, hours to be predicted, size of the forecast area, and whether or not CUDA was used. The results were the distribution of the replicate data and the significance of the resulting data. Discussion of the results: In the analysis of the results a significant gain was observed, making use of CUDA integrated to the model, although in some cases a performance worsening occurs, however, the tests performed occur in computational environments not suitable for experiments of high applications performance, and may therefore harm the results generated. Conclusions: It can be stated that this dissertation contributed to the improvement of the Eta model. The knowledge obtained will serve as a basis for new work involving the model. It is emphasized that due to the unavailability of an optimized computing environment for highperformance applications, new tests and the benefits that can be obtained using CUDA technology in conjunction with MPI in the Eta model of weather are required.